linYD0718/open-perfectblend-qwen3-4b-nonthinking

Open PerfectBlend Qwen3-4B Non-Thinking

This dataset contains 1,349,812 training conversations derived from `mlabonne/open-perfectblend`. Each assistant turn was regenerated sequentially with `Qwen/Qwen3-4B`, conditioned on the preceding conversation, with thinking disabled.

Data preparation

Empty or otherwise invalid source conversations were removed before a deterministic train/evaluation split. The split used seed 42 and a held-out fraction of 0.05. Only the 1,349,812-row training split is included here; the 71,043-row held-out split is not included.

Generation parameters:

• —temperature: 0.7
• —top-p: 0.8
• —top-k: 20
• —min-p: 0.0
• —maximum new tokens per assistant turn: 4,096
• —thinking: disabled

Schema

Each line in train.jsonl is a JSON object with:

• —id: source sample identifier
• —conversations: ordered messages with role and content
• —status: success

All 1,349,812 records were validated as parseable JSON with unique IDs, valid conversation roles, non-empty assistant responses, and no thinking fields.

Artifact integrity

• —File: train.jsonl
• —Records: 1,349,812
• —Size: 5,827,170,794 bytes
• —SHA-256: 3bb311b2c9d48939c93a1eabdddd0c8b78c26a3abfd4274ebd7d3a605ff5daa7

Compact token cache

A compact DeepSpec token-cache snapshot is available under token_cache/. It stores the fully rendered input_ids as int32 and the aligned assistant loss_mask as uint8; it does not contain target hidden states or an attention KV cache.

• —Source rows processed: 1,349,812
• —Valid cached samples: 1,339,718
• —Samples filtered below 14 assistant loss tokens: 10,094
• —Immutable segments: 14
• —Serialized tokens: 1,345,413,651
• —Binary token/mask payload: 6,727,068,255 bytes
• —Binary index payload: 53,588,720 bytes
• —Chat template: qwen
• —Maximum rendered length: 4,096 tokens

The uploaded manifests use the public model identifier Qwen/Qwen3-4B and the repository-relative source name train.jsonl; private preprocessing paths and source-index files are excluded. See token_cache/README.md, token_cache/cache_metadata.json, and token_cache/checksums.sha256 for layout, provenance, and integrity information.

License and attribution

The source dataset and Qwen3-4B model are distributed under the Apache License 2.0. This derived dataset is released under Apache-2.0. Users are responsible for reviewing generated content and complying with applicable source-model and dataset terms.